Created
July 26, 2023 21:47
-
-
Save emchristiansen/db80f5e85c791f6bb5bba5b78b750cd9 to your computer and use it in GitHub Desktop.
Proposed gradient tape design pattern
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
use color_eyre::Result; | |
use dfdx::shapes::Rank0; | |
use dfdx::tensor::AsArray; | |
use dfdx::tensor::Cpu; | |
use dfdx::tensor::NoneTape; | |
use dfdx::tensor::OwnedTape; | |
use dfdx::tensor::PutTape; | |
use dfdx::tensor::SplitTape; | |
use dfdx::tensor::Tape; | |
use dfdx::tensor::Tensor; | |
use dfdx::tensor::TensorFrom; | |
use dfdx::tensor::Trace; | |
use dfdx::tensor::WithEmptyTape; | |
use dfdx::tensor_ops::Backward; | |
use dfdx::tensor_ops::TryStack; | |
type Tensor_ = Tensor<Rank0, f32, Cpu, NoneTape>; | |
// Another design for forward fns. | |
// The status quo is to associate the tape with the inputs and outputs, and | |
// you have to remember which input has the tape and which output should have | |
// it. | |
// Instead, why not just pass it in and out explicitly? | |
fn forward<T>( | |
tape: T, | |
x: Tensor_, | |
y: Tensor_, | |
) -> ( | |
T, | |
Tensor_, | |
Tensor_, | |
) | |
where | |
T: Tape<f32, Cpu>, | |
{ | |
let (prod_1, tape) = (x | |
.clone() | |
.put_tape(tape) | |
* y.clone()) | |
.split_tape(); | |
let (prod_2, tape) = (x.put_tape(tape) * y * 2.0).split_tape(); | |
( | |
tape, prod_1, prod_2, | |
) | |
} | |
fn main() -> Result<()> | |
{ | |
let dev = Cpu::default(); | |
let x = dev.tensor(1.0); | |
let y = dev.tensor(2.0); | |
let (tape, out0, out1) = forward( | |
OwnedTape::default(), | |
x.clone(), | |
y.clone(), | |
); | |
// Now we can decide at the call-site which gradient we want. | |
let grad = out1 | |
.put_tape(tape) | |
.backward(); | |
dbg!(&grad | |
.get(&x) | |
.array()); | |
dbg!(&grad | |
.get(&y) | |
.array()); | |
Ok(()) | |
} |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment